Re-run #31's unlock probe on 17.4.0 — its criterion may already be met and nobody has looked

[zhuangjianguo]

⛔ Observation only. This card produces NO code and NO pull request. Its deliverable is a measurement and a report. ⛔ Do not fix anything you find, do not add a field, do not touch DESIGN.md, and ⛔ do not change #31's labels — the seat does that after reading your report.

Why now

Decision #31 is pm:blocked, and it states its own unlock criterion verbatim in its header:

解锁判据是消费方可安装（本仓升到含该修复的版本后那条错路确实报错），⛔ 不是「上游 PR 已合并」。

This repository upgraded to @objectstack/* 17.4.0 this session — PR #37, merged as 5d5a254. Nobody has re-run the probe since. A blocked card whose blocker may have cleared is the failure mode that keeps a decision parked forever, because the condition changes silently and no event fires.

The comment block in src/datasets/cycle-time.dataset.ts makes the same gap visible from the other side: it names its measurement's versions honestly — 「Measured on @objectstack/spec 17.3.0 + @objectstack/service-analytics 17.3.0」 — so a 17.4.0 decision is currently resting on a 17.3.0 reading. ⚠️ That the comment names its versions is correct practice and is the only reason this is visible at all; ⛔ it is not a defect and is not to be "fixed".

The probe, which #31 already specifies

#31 records the exact wrong path and what it did on 17.3.0:

derived: { op: 'difference', of: [avg_activated, avg_submitted] }
  → parsed, ran, rendered  -0.849999999999909

and why:

sqlite> select typeof(submitted_at), submitted_at from clm_contract limit 1;
text|2026-05-19T00:00:00.000Z
sqlite> select avg(submitted_at) from clm_contract;
2025.9166666666667          ← the average YEAR, not the average date

Establish, on 17.4.0, whether that path now errors. Concretely:

1. Does derived: { op: 'difference', … } over two datetime averages still parse, run and render a number — or does it now refuse, and at which layer (author-time validate, query build, or execution)?
2. Does AVG() over a Field.datetime still return an average year, or does the analytics layer now refuse the aggregate?
3. Whichever answer: what exactly does the operator see? An exception, a validation message, an empty widget, or a number. [Decision] §09 要「各段时长」和「超 SLA」，但语义层算不出时间差——两问共用一个补法 #31's whole argument is that the dangerous state is a clean, plausible, wrong number, so "it errors" and "it renders nothing" are different outcomes and both differ from "it still lies".

Use a scratch dataset in your own worktree. ⛔ Do not commit it — restore and prove the tree byte-identical by state (git diff HEAD empty, git hash-object matching git rev-parse HEAD:<path>), the way PR #70, #72, #74 and #79 did.

What the answer means — ⛔ do not decide this yourself

• If the wrong path now errors: [Decision] §09 要「各段时长」和「超 SLA」，但语义层算不出时间差——两问共用一个补法 #31's criterion is met as written. ⚠️ Note that this does not deliver the metric — [Decision] §09 要「各段时长」和「超 SLA」，但语义层算不出时间差——两问共用一个补法 #31's own public correction says the upstream fix (objectstack-ai/objectstack#16778) chose to refuse rather than to provide, so durations stay uncomputable. What changes is that the trap is closed. Report it; the seat decides what [Decision] §09 要「各段时长」和「超 SLA」，但语义层算不出时间差——两问共用一个补法 #31 becomes.
• If it still renders a number: the criterion is not met, [Decision] §09 要「各段时长」和「超 SLA」，但语义层算不出时间差——两问共用一个补法 #31 stays pm:blocked, and you have refreshed a stale measurement — which is worth the run on its own, since the current comment cites 17.3.0.
• If the answer is neither — a different refusal, a different number, a changed shape — report exactly what you saw. ⛔ Do not force it into one of the two boxes above.

Constraints

• ⛔ No code, no PR, no DESIGN.md, no label changes. Report as a comment on this issue.
• ⛔ Do not re-open the 1A/2A vs 1B/2B question. That was adjudicated 1C by the maintainer on 2026-09-09 and §09 already records the outcome (PR Record four maintainer rulings in DESIGN.md §03, §06 and §09 #40, 1127e52). You are measuring one fact, not revisiting a ruling.
• ⛔ Platform gaps: report, never patch. If you find a new platform defect, say so — the seat routes it upstream.
• Name the exact versions you measured on, the way the existing comment does. That practice is why this card exists.

Acceptance

• A direct answer to each of the three questions, with the reading that produced it — the request or config that was run, and the raw output. ⛔ Not a summary of what you expect.
• The versions installed at measurement time, read from the tree rather than assumed.
• Whether the tree is byte-identical afterwards, proved by state.
• If the probe cannot be run at all, say why in the same detail — an unrunnable probe is a finding, not a failure.

Provenance

Noticed by the PM seat while reviewing PR #79 (#75), which quotes #31's unlock criterion inside a comment it was reframing. The dev on #75 correctly did not chase it — it was out of that card's scope. ⚠️ The seat should have noticed at the 17.4.0 upgrade (PR #37) that a blocked card's condition might have moved; it did not, and this card is that catch-up.

Related

#31 (pm:blocked, the card this is for) · #37 / 5d5a254 (the 17.4.0 upgrade) · objectstack-ai/objectstack#16737 (the blocker) · objectstack-ai/objectstack#16778 (the in-flight fix that refuses rather than provides) · #75 / PR #79 (where this surfaced) · src/datasets/cycle-time.dataset.ts (the 17.3.0 measurement)

[zhuangjianguo]

Claim: session session_01R3n3GGzobdegM4HUzah1iR · branch claude/issue-80-recheck-31-unlock (a branch is claimed for isolation only — ⛔ this card pushes nothing)

PM dispatch. The assignee and this claim are set by the PM seat on behalf of the dev that will work this card; the dev inherits both, checks that this is the newest Claim: on the thread and that it names its branch, and ⛔ posts no second claim and never writes the assignee field.

Base off current origin/main (96683f7).

⛔ This card has no deliverable in the repository. No commit, no push, no PR. Its output is a comment on this issue. The branch exists so your worktree is isolated and the scratch dataset you need has somewhere to live; it is ⛔ never pushed.

Notes on top of the card body:

• The single question is whether [Decision] §09 要「各段时长」和「超 SLA」，但语义层算不出时间差——两问共用一个补法 #31's own criterion is met, and [Decision] §09 要「各段时长」和「超 SLA」，但语义层算不出时间差——两问共用一个补法 #31 wrote it: 「本仓升到含该修复的版本后那条错路确实报错」. Not "did the upstream PR merge" — [Decision] §09 要「各段时长」和「超 SLA」，但语义层算不出时间差——两问共用一个补法 #31 rules that out explicitly. So the reading has to come from this repo, on the version it actually has installed, running the wrong path.
• Read the installed versions from the tree, not from package.json's range and not from the card. package.json carries ^17.0.0; what is installed is what matters, and PR Upgrade the @objectstack/* dependencies from 17.3.0 to 17.4.0 #37 moved it. That number goes in your report.
• Distinguish three outcomes, and ⛔ do not collapse them: it errors · it renders nothing · it still renders a plausible number. [Decision] §09 要「各段时长」和「超 SLA」，但语义层算不出时间差——两问共用一个补法 #31's entire argument is that the third is the dangerous state, so "no longer works" is not a sufficient answer — say which of the three, at which layer, and quote what the operator sees.
• A negative is a result. If it still renders -0.85-shaped output, that is the answer and it refreshes a measurement currently citing 17.3.0. ⛔ Do not treat "the criterion is not met" as a failed run.
• ⛔ Do not decide what happens to [Decision] §09 要「各段时长」和「超 SLA」，但语义层算不出时间差——两问共用一个补法 #31. Report the fact; the seat reads it and decides, and the 1C ruling itself is settled and ⛔ not reopenable by you.
• If the probe cannot be run at all — the dataset shape is refused at author time, the demo will not boot, whatever — that is a finding. Report it with the same detail you would report a success. ⛔ Do not substitute reasoning about what 17.4.0 probably does.

Acceptance

• Each of the card's three questions answered from a reading, with the config or request you ran and its raw output. ⛔ No summaries of expectations.
• Installed @objectstack/* versions, read from the tree.
• The tree byte-identical when you finish, proved by state (git diff HEAD empty; git hash-object matching git rev-parse HEAD:<path> for anything you touched) — ⛔ never by a command's exit code. PR Five comments in three files still say DESIGN.md §09 asks for the metrics its 2026-09-09 ruling placed outside V1.0 #79's two-leg ablation is the shape.
• ⛔ Nothing committed, nothing pushed, no PR opened.

---

Generated by Claude Code

[claude]

os-dev-report

{
  "issue": 80,
  "status": "done",
  "branch": "claude/issue-80-recheck-31-unlock",
  "pr": null,
  "premise_still_valid": true,
  "summary": "Re-ran decision #31's unlock probe on the platform version this repo actually has installed (17.4.0 across all 54 @objectstack packages, read from node_modules). CRITERION NOT MET. The wrong path still parses, still runs, and still renders a clean plausible number at every one of the three layers: pnpm validate / lint / typecheck all exit 0 with the derived measure landing verbatim in dist/objectstack.json; POST /api/v1/analytics/dataset/query answers HTTP 200; and the dashboard tile renders -0.86 under the label 'Avg Cycle (days?)' with no banner, no empty state and no message. Of #31's three outcomes this is the THIRD and most dangerous one — it still lies. AVG() over a Field.datetime still returns the average YEAR (2025.8833333333334) and no layer refuses the aggregate. 17.4.0 shipped two SIBLINGS of the blocker's family (#15768 min/max described as time, #16098 min/max described as string) but NOT the refusal itself: no installed CHANGELOG mentions 16737 or 16778, and service-analytics 17.4.0's own release note states the measured state verbatim — 'sum / avg over a temporal column are refused by no layer and answered by the backend (an epoch mean on SQLite, an error on Postgres)'. This also refreshes the 17.3.0 measurement quoted in src/datasets/cycle-time.dataset.ts: same shape, new numbers. ⛔ No code, no commit, no push, no PR, no label change; tree proved byte-identical by state.",
  "tests": "VERSIONS, read from the tree (not from package.json's ^17.0.0 range, not from the card): node_modules/@objectstack/{spec,runtime,cli,plugin-hono-server}/package.json all report 17.4.0; node_modules/.pnpm holds @objectstack+service-analytics@17.4.0 and 53 further @objectstack packages, every one at 17.4.0; better-sqlite3@13.0.3. Worktree /home/user/hotclm-issue-80 off origin/main 96683f7. || Q1 — DOES THE DERIVED DIFFERENCE STILL PARSE, RUN AND RENDER? YES, at all three layers, nothing refuses. (a) AUTHOR TIME: a scratch dataset probe31_cycle on clm_contract carrying { name: 'avg_submitted', aggregate: 'avg', field: 'submitted_at' }, { name: 'avg_activated', aggregate: 'avg', field: 'activated_at' } and { name: 'cycle_days', derived: { op: 'difference', of: ['avg_activated','avg_submitted'] } }. `pnpm validate` exit 0, printing '✓ Validation passed (923ms)'; `pnpm lint` exit 0 (its only probe-related output is 8 i18n/missing-dataset and 2 i18n/missing-widget zh-CN warnings — nothing about the aggregate); `pnpm typecheck` exit 0. Not a phantom green: `pnpm build` exit 0 and dist/objectstack.json lists datasets ['contract_metrics','contract_cycle_time','obligation_metrics','payment_metrics','probe31_cycle'] with cycle_days present verbatim as {\"name\":\"cycle_days\",\"label\":\"Avg Cycle (days?)\",\"derived\":{\"op\":\"difference\",\"of\":[\"avg_activated\",\"avg_submitted\"]}}. Read against the source: @objectstack/spec 17.4.0 still declares DerivedMeasureOp = z.enum(['ratio','sum','difference','product']) in src/ui/dataset.zod.ts and its superRefine checks only self-reference and unknown-measure-name — no aggregate/field-type pairing check exists. (b) QUERY BUILD: POST /api/v1/analytics/dataset/query, body {\"datasetName\":\"probe31_cycle\",\"selection\":{\"measures\":[\"contract_count\",\"avg_submitted\",\"avg_activated\",\"cycle_days\"]}} answered HTTP_CODE=200, and the response's own `sql` echo is 'SELECT COUNT(*) AS \"contract_count\", AVG(submitted_at) AS \"avg_submitted\", AVG(activated_at) AS \"avg_activated\" FROM \"clm_contract\"' — the aggregate is emitted, not declined. (c) EXECUTION: rows [{\"contract_count\":120,\"avg_submitted\":2025.8833333333334,\"avg_activated\":2025.0277777777778,\"cycle_days\":-0.8555555555556111}], fields all {\"type\":\"number\"}. || Q2 — DOES AVG() OVER A Field.datetime STILL RETURN AN AVERAGE YEAR? YES, and no layer refuses it. Raw better-sqlite3 13.0.3 readings against the demo database .objectstack/data/objectstack.db, the same statements #31 ran: `select typeof(submitted_at), submitted_at from clm_contract where submitted_at is not null limit 1` → [{\"t\":\"text\",\"submitted_at\":\"2026-08-16T00:00:00.000Z\"}]; `select avg(submitted_at) from clm_contract` → [{\"avg_submitted\":2025.8833333333334}]; `select avg(activated_at) from clm_contract` → [{\"avg_activated\":2025.0277777777778}]; `select avg(activated_at) - avg(submitted_at) from clm_contract` → [{\"difference\":-0.8555555555556111}] — byte-for-byte the value the analytics layer returned for cycle_days, which proves the derived op runs on the coerced YEARS. `select count(*), count(submitted_at), count(activated_at) from clm_contract` → [{\"n_rows\":120,\"n_submitted\":60,\"n_activated\":72}]. The DDL declares `submitted_at datetime` / `activated_at datetime`; SQLite is dynamically typed and stores the ISO string, so typeof is 'text' — unchanged from 17.3.0. The analytics layer describes the column as {\"name\":\"avg_submitted\",\"type\":\"number\"}, which is not a mislabel: 2025.88 IS a number. What is missing is any refusal. || Q3 — WHAT EXACTLY DOES THE OPERATOR SEE? A NUMBER. Not an exception, not a validation message, not an empty widget. Chromium 1194 (/opt/pw-browsers/chromium-1194/chrome-linux/chrome), signed in as admin@objectos.ai at http://localhost:3480/_console/apps/clm/dashboard/legal_workbench, two scratch metric tiles added to legal_workbench. Rendered text, quoted from the live page: tile 'PROBE derived difference' shows the value '-0.86' over the sub-label 'Avg Cycle (days?)'; tile 'PROBE avg(submitted_at)' shows '2,025.88' over 'Avg Submitted' — thousands-separated, so it reads like a plausible count. Both sit in the normal metric-tile chrome beside the board's real tiles (Awaiting Intake 6, In Review 12, In Review Over 30 Days 6, Waiting on Counterparty 2, Approved This Month 1, Pipeline by Stage), with no error banner, no empty state and no console error from the widgets. The browser's own network capture for the derived tile: request {\"datasetName\":\"probe31_cycle\",\"selection\":{\"dimensions\":[],\"measures\":[\"cycle_days\"]}} → status 200, body {\"rows\":[{\"avg_activated\":2025.0277777777778,\"avg_submitted\":2025.8833333333334,\"cycle_days\":-0.8555555555556111}],...}. || WHY THE NUMBER DIFFERS FROM #31's: #31 read -0.849999999999909, this run reads -0.8555555555556111. Same shape, same mechanism, same row counts (60 submitted / 72 activated); the demo fixture's dates are relative to the boot date, so the two average years shift. Nothing about the behaviour changed. || BOOT HEALTH: 38 plugins loaded including AnalyticsServicePlugin; '✓ Server is ready'; the only boot warning is the known Seeder inline-budget one plus the documented first-boot missing-owner ERROR lines this repo's own demo note explains. || RESTORATION, proved by state and never by an exit code: `git status --porcelain` before restore listed ' M src/dashboards/legal.dashboard.ts', ' M src/datasets/index.ts', '?? src/datasets/probe31.dataset.ts'. Restored with `git checkout HEAD -- ...` (pointing at HEAD, not the index) plus `rm -f` of the untracked probe. After: `git diff HEAD` empty, `git status --porcelain` empty; git hash-object src/dashboards/legal.dashboard.ts = 7ee7356a9063df2c2c67f4e6634e229abf0760c6 = git rev-parse HEAD:src/dashboards/legal.dashboard.ts (MATCH); git hash-object src/datasets/index.ts = 14ce0601805306c73537a3b724ece7451ea7b67f = git rev-parse HEAD:src/datasets/index.ts (MATCH); src/datasets/probe31.dataset.ts absent. HEAD still 96683f7, nothing committed; `git ls-remote origin claude/issue-80-recheck-31-unlock` finds nothing — the branch was never pushed. Demo server stopped by the PID/PGID I recorded (server PID 8693, PGID 8575), never by name; port 3480 free afterwards.",
  "mcp_calls": "0 — every GitHub read and this write went through repo-scoped REST from this container; no MCP GitHub tool was called",
  "open_questions": [],
  "out_of_scope_findings": [
    "noted, not filed: 17.4.0 shipped two siblings of the blocker's family (#15768 min/max described as `time`, #16098 min/max described as `string`) but not the refusal — no installed CHANGELOG under node_modules mentions 16737 or 16778. A reader skimming service-analytics 17.4.0's temporal release notes could mistake the sibling for the fix. Successor: the seat, when it next reads #16737's status.",
    "noted, not filed: service-analytics 17.4.0's own release note for #15768 states the measured state verbatim — 'sum / avg over a temporal column are refused by no layer and answered by the backend (an epoch mean on SQLite, an error on Postgres), so there is no single value for a type to describe and none is invented'. That is the platform declaring this hole open on the exact version this repo runs, which is the strongest single citation for #31 staying blocked. Successor: the seat.",
    "noted, not filed: selecting ONLY the derived measure returns three columns, not one — the response for {\"measures\":[\"cycle_days\"]} carries avg_activated and avg_submitted beside it. Harmless on a metric tile (it reads the named measure) and arguably necessary since the derived op needs its operands, but a table-shaped widget selecting one derived measure would receive its operand columns too. Observation, not a defect, and NOT a new platform gap. Successor: none.",
    "noted, not filed: the 17.3.0 measurement quoted in src/datasets/cycle-time.dataset.ts is now refreshed by this run — same shape, new numbers, same conclusion. The card is explicit that the comment naming its versions is correct practice and ⛔ not to be 'fixed', so nothing is proposed here; the refreshed reading is recorded in this comment for whoever next edits that file. Successor: whichever card next touches that comment."
  ]
}

---

Appendix — raw output, in the order it was taken

Versions, from the tree (node_modules, not package.json):

@objectstack/spec               17.4.0
@objectstack/runtime            17.4.0
@objectstack/cli                17.4.0
@objectstack/plugin-hono-server 17.4.0
@objectstack/service-analytics  17.4.0     (node_modules/.pnpm/@objectstack+service-analytics@17.4.0)
better-sqlite3                  13.0.3
… 54 @objectstack packages in the store, all 17.4.0, zero 17.3.0 residue

Layer 1 — author time. The scratch dataset (never committed):

export const Probe31Dataset = defineDataset({
  name: 'probe31_cycle',
  object: 'clm_contract',
  dimensions: [{ name: 'status', label: 'Status', field: 'status', type: 'string' }],
  measures: [
    { name: 'contract_count', label: 'Contracts', aggregate: 'count' },
    { name: 'avg_submitted',  label: 'Avg Submitted',  aggregate: 'avg', field: 'submitted_at' },
    { name: 'avg_activated',  label: 'Avg Activated',  aggregate: 'avg', field: 'activated_at' },
    { name: 'cycle_days', label: 'Avg Cycle (days?)',
      derived: { op: 'difference', of: ['avg_activated', 'avg_submitted'] } },
  ],
});

VALIDATE_EXIT=0     ✓ Validation passed (923ms)
LINT_EXIT=0         (probe-related output: 8 × i18n/missing-dataset, 2 × i18n/missing-widget — zh-CN keys only)
TYPECHECK_EXIT=0
BUILD_EXIT=0        datasets registered: ['contract_metrics','contract_cycle_time','obligation_metrics','payment_metrics','probe31_cycle']

Layer 2/3 — query build and execution.

$ curl -X POST http://localhost:3480/api/v1/analytics/dataset/query \
    -d '{"datasetName":"probe31_cycle","selection":{"measures":["contract_count","avg_submitted","avg_activated","cycle_days"]}}'
HTTP_CODE=200
{
    "rows": [
        {
            "contract_count": 120,
            "avg_submitted": 2025.8833333333334,
            "avg_activated": 2025.0277777777778,
            "cycle_days": -0.8555555555556111
        }
    ],
    "fields": [
        { "name": "contract_count", "type": "number", "label": "Contracts" },
        { "name": "avg_submitted",  "type": "number", "label": "Avg Submitted" },
        { "name": "avg_activated",  "type": "number", "label": "Avg Activated" },
        { "name": "cycle_days",     "type": "number", "label": "Avg Cycle (days?)" }
    ],
    "sql": "SELECT COUNT(*) AS \"contract_count\", AVG(submitted_at) AS \"avg_submitted\", AVG(activated_at) AS \"avg_activated\" FROM \"clm_contract\""
}

Storage — the same statements #31 ran, on 17.4.0:

sqlite> select typeof(submitted_at) as t, submitted_at from clm_contract where submitted_at is not null limit 1
[{"t":"text","submitted_at":"2026-08-16T00:00:00.000Z"}]
sqlite> select typeof(activated_at) as t, activated_at from clm_contract where activated_at is not null limit 1
[{"t":"text","activated_at":"2026-05-18T00:00:00.000Z"}]
sqlite> select avg(submitted_at) as avg_submitted from clm_contract
[{"avg_submitted":2025.8833333333334}]
sqlite> select avg(activated_at) as avg_activated from clm_contract
[{"avg_activated":2025.0277777777778}]
sqlite> select avg(activated_at) - avg(submitted_at) as difference from clm_contract
[{"difference":-0.8555555555556111}]
sqlite> select count(*) as n_rows, count(submitted_at) as n_submitted, count(activated_at) as n_activated from clm_contract
[{"n_rows":120,"n_submitted":60,"n_activated":72}]

clm_contract DDL, the two columns: `submitted_at` datetime, … `activated_at` datetime — declared datetime, stored as text, so AVG() still takes SQLite's text→numeric coercion of 2026-… and answers the year.

What the operator sees — Legal Workbench, Chromium, signed in as the dev admin:

PROBE avg(submitted_at)          PROBE derived difference
2,025.88                         -0.86
Avg Submitted                    Avg Cycle (days?)

Two ordinary metric tiles. No banner, no empty state, no validation message, no widget console error. The browser's own request/response for the right-hand tile:

request  {"datasetName":"probe31_cycle","selection":{"dimensions":[],"measures":["cycle_days"]}}
status   200
body     {"rows":[{"avg_activated":2025.0277777777778,"avg_submitted":2025.8833333333334,
          "cycle_days":-0.8555555555556111}], …,
          "sql":"SELECT AVG(activated_at) AS \"avg_activated\", AVG(submitted_at) AS \"avg_submitted\" FROM \"clm_contract\""}

Verdict against #31's own criterion — 「本仓升到含该修复的版本后那条错路确实报错」: on 17.4.0 that path does not error. Of the three outcomes the card refuses to let collapse — it errors · it renders nothing · it still renders a plausible number — this run measured the third. ⛔ What #31 becomes is the seat's call, not mine.

---

Generated by Claude Code